Model training optimization system based on classical-quantum hybrid architecture

Abstract

The invention relates to model training, in particular to a model training optimization system based on a classical-quantum hybrid architecture, which comprises a system configuration module used for configuring and scheduling resources required by a data preprocessing and quantum state coding module and a quantum circuit optimization and error mitigation module; the data preprocessing and quantum state coding module is used for preprocessing original data by adopting classic computing equipment, carrying out quantum state coding on the preprocessed data by adopting quantum computing equipment, and embedding a quantum state into a high-dimensional quantum feature space; the quantum circuit optimization and error mitigation module is used for performing calculation optimization and result optimization on the preprocessed data and quantum states, and finally outputting an optimal result in combination with a classical calculation result and a quantum calculation result; according to the technical scheme provided by the invention, the defect that efficient model training is difficult to promote by constructing a classical-quantum hybrid architecture in the prior art can be effectively overcome.

Description

Model training optimization system based on classical-quantum hybrid architecture Technical Field

[0001] This invention relates to model training, and more specifically to a model training optimization system based on a classical-quantum hybrid architecture. Background Technology

[0002] Quantum computing utilizes the properties of quantum superposition and quantum entanglement, and theoretically can achieve exponential speedups relative to classical computers on specific computing tasks. Quantum computing research and technological innovation are developing rapidly and have become a research hotspot in the field of cutting-edge technology.

[0003] Quantum machine learning and large-scale model training, as an emerging interdisciplinary field, aim to leverage the advantages of quantum computing to address the computational bottlenecks in classical machine learning. Quantum machine learning utilizes the parallelism and efficiency of quantum computing to accelerate machine learning tasks, process more complex datasets in less time, and improve model training speed and prediction accuracy. It shows particular potential in handling high-dimensional data and complex pattern recognition.

[0004] Despite the promising future of quantum machine learning, current technologies still have many limitations:

[0005] 1) In terms of technological maturity: Quantum computing is in a critical period of technological breakthroughs. There are still many challenges in terms of the stability, scalability, control precision and error correction capability of qubits. Current quantum computing prototypes cannot meet the scale and precision requirements of fault-tolerant general-purpose quantum computing, which directly limits the application effect of quantum machine learning in real-world scenarios.

[0006] 2) In terms of algorithms and software ecosystem: Quantum computing software, as a key bridge connecting quantum computing hardware and practical applications, is still in its early stages. Software tools are showing obvious fragmentation and diversification. At the same time, efficient quantum algorithms for complex scenarios still need to be continuously explored and iterated.

[0007] In patent application CN117151204A, a model training method, data processing method, and apparatus based on quantum computing are disclosed. This scheme uses a quantum neural network model composed of quantum circuits, which theoretically can improve model training efficiency. Combining world model prediction and reinforcement learning, it is suitable for decision-making problems in unknown environments. However, this scheme is specifically designed for reinforcement learning scenarios, resulting in a narrow application scope. Furthermore, the quantum circuit design is relatively fixed, lacking adaptive quantum circuit optimization strategies, and it does not fully consider the noise problems of current quantum computing devices and corresponding error mitigation mechanisms.

[0008] In patent application CN117556908A, a quantum computing-assisted language large-scale model training system and method are disclosed. This solution transforms the training of large language models, which was originally based on classical computation, into a state that can be recognized by quantum computing, thus initially realizing a quantum computing replacement for existing large language model training and possessing a certain degree of technological foresight. However, this solution is specifically designed for quantum computing replacement of large language model training and lacks universality applicable to all model training scenarios. Furthermore, it does not fully consider that quantum computing technology is still in a stage of rapid development, and its stability and availability may lead to ineffective transfer work and the risk of repetitive labor in adapting to technological development. Therefore, it lacks the practical feasibility of using classical computation as the primary method and quantum computing as a supplement based on its own advantages. Summary of the Invention

[0009] (a) Technical problems to be solved

[0010] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a model training optimization system based on classical-quantum hybrid architecture, which can effectively overcome the shortcomings of the existing technology in that it is difficult to promote efficient model training by constructing classical-quantum hybrid architecture.

[0011] (II) Technical Solution

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] A model training and optimization system based on a classical-quantum hybrid architecture includes a system configuration module, a data preprocessing and quantum state encoding module, and a quantum circuit optimization and error mitigation module.

[0014] The system configuration module is used to configure and schedule the resources required by the data preprocessing and quantum state encoding module and the quantum circuit optimization and error mitigation module;

[0015] The data preprocessing and quantum state encoding module uses classical computing devices to preprocess the raw data, and uses quantum computing devices to encode the preprocessed data into quantum states, embedding the quantum states into a high-dimensional quantum feature space.

[0016] The quantum circuit optimization and error mitigation module performs computational optimization and result optimization on the preprocessed data and quantum states, and finally combines the classical calculation results and quantum calculation results to output the optimal result.

[0017] Preferably, the system configuration module includes a task scheduling layer, a computation execution layer, and a resource management layer;

[0018] The task scheduling layer is deployed on high-performance classical computing devices, using multi-core CPUs, GPU clusters, or distributed computing nodes. The task parsing unit uses natural language processing technology to parse user-submitted training tasks, identifying task type, data characteristics, and performance requirements. The resource allocation unit dynamically allocates classical and quantum computing resources based on task requirements and system resource status. The process scheduling unit uses directed acyclic graph (DAG) technology to manage the task execution process, ensuring the correct execution of dependencies between tasks.

[0019] The computation execution layer employs a heterogeneous computing architecture and intelligent decomposition algorithm. Based on the characteristics of the computation task, the load of classical computing devices, and the performance of quantum computing devices, it dynamically determines whether a subtask is suitable for classical processing or quantum acceleration. The classical computing devices include CPUs and GPUs, which are responsible for executing traditional computing tasks. The quantum computing devices are accessed through a quantum cloud platform or a local quantum processor, and support various quantum hardware computing platforms, including superconducting qubits and ion traps, through standard computing power access interfaces.

[0020] The resource management layer comprises a classical resource management unit and a quantum resource management unit. The classical resource management unit is responsible for managing classical computing resources, storage resources, and network resources, and uses distributed resource management technology to achieve unified management of large-scale classical computing resources. The quantum resource management unit is responsible for managing qubit resources, quantum gate resources, and quantum circuit compilation optimization, and uses quantum circuit compilation optimization algorithms to transform high-level quantum circuit descriptions into low-level instructions executable by quantum computing devices.

[0021] Preferably, the data preprocessing and quantum state encoding module includes a data preprocessing unit, a quantum state encoding unit, and a quantum feature mapping unit;

[0022] The data preprocessing unit uses classical computing equipment to preprocess the raw data;

[0023] The quantum state encoding unit uses a quantum state encoder to encode the preprocessed data into quantum states.

[0024] The quantum feature mapping unit embeds the quantum state into a high-dimensional quantum feature space after the data is encoded into a quantum state using quantum feature mapping technology.

[0025] Preferably, the data preprocessing unit uses classical computing equipment to preprocess the raw data, including:

[0026] Data cleaning: handling missing values, outliers, and duplicate values ​​to ensure data quality;

[0027] Feature engineering: Enhances feature representation capabilities by selecting relevant features, constructing new features, and transforming feature representations;

[0028] Data normalization: scaling data to a range suitable for quantum state encoding;

[0029] Dimensional Adjustment: The feature dimension is adjusted according to the number of qubits, and principal component analysis or an autoencoder is used to compress the feature dimension to 2. n Dimension, n is the number of qubits.

[0030] Preferably, the quantum state encoding unit uses a quantum state encoder to perform quantum state encoding on the preprocessed data, including:

[0031] An adaptive encoding method selection mechanism is introduced to automatically select the most suitable encoding method from the following options based on task requirements and data characteristics, thereby improving encoding efficiency:

[0032] Basis encoding: directly mapping the binary representation of classical data to the quantum ground state;

[0033] Amplitude encoding: Representing classical data as the amplitude of quantum states to achieve exponentially compressed storage;

[0034] Angle encoding: Maps classical data to the rotation angle of a quantum gate, suitable for continuous data;

[0035] Hamiltonian encoding: Encoding classical data into Hamiltonian parameters for use in quantum simulation tasks.

[0036] Preferably, after the data is encoded into a quantum state, the quantum feature mapping unit uses quantum feature mapping technology to embed the quantum state into a high-dimensional quantum feature space, including:

[0037] Quantum states are embedded into high-dimensional quantum feature spaces using custom quantum circuits or parameterized quantum circuits:

[0038] Custom quantum circuits: Achieving complex feature mappings by designing quantum circuit structures;

[0039] Parameterized quantum circuits: Using trainable parameterized quantum gate sequences, feature mappings are optimized by adjusting parameters.

[0040] Preferably, the quantum circuit optimization and error mitigation module includes an adaptive quantum circuit optimization unit and a quantum error mitigation unit;

[0041] The adaptive quantum circuit optimization unit adopts an adaptive quantum circuit optimization strategy to optimize quantum circuits based on task requirements and hardware characteristics.

[0042] The quantum error mitigation unit employs a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing.

[0043] Preferably, the adaptive quantum circuit optimization unit employs an adaptive quantum circuit optimization strategy to optimize the quantum circuit according to task requirements and hardware characteristics, including:

[0044] Quantum circuit structure optimization: Select suitable circuit templates from the parameterized quantum circuit template library according to task requirements to balance circuit expressiveness and complexity;

[0045] Dynamic circuit pruning: By analyzing the importance of each quantum gate in a quantum circuit, redundant quantum gates that have little impact on quantum computing results are removed, thereby reducing the circuit depth;

[0046] Noise-adaptive compilation: Based on the noise characteristics and error rate of the quantum computing device, select a quantum gate decomposition scheme and a qubit mapping strategy with lower noise.

[0047] Preferably, the quantum error mitigation unit employs a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing, including:

[0048] Zero-noise extrapolation: The original quantum circuit is randomly compiled into a series of equivalent quantum circuits using random compilation technology. The noise level is adjusted using dynamic decoupling technology, and these quantum circuits are run under different noise levels. The quantum computing results at each noise level are collected, and the quantum computing results at zero noise are estimated using an extrapolation algorithm.

[0049] Measurement error mitigation: Construct a measurement error calibration matrix, measure the output distribution under all calculated ground-state inputs, and correct the actual measurement results using deconvolution or maximum likelihood estimation;

[0050] Probabilistic error elimination: The original quantum circuit is randomly compiled into a series of equivalent quantum circuits using random compilation technology. With the help of the quasi-probabilistic decomposition framework, the noiseless quantum operation is approximately expanded into a convex combination of noisy quantum operations under a specific noise model. The impact of noise on quantum information processing is analyzed, and the impact of noise is offset by weighted averaging of multiple experimental results.

[0051] (III) Beneficial Effects

[0052] Compared with existing technologies, the model training optimization system based on a classical-quantum hybrid architecture provided by this invention solves several key problems in existing quantum machine learning technologies by constructing a classical-quantum deep collaborative architecture and an adaptive quantum circuit optimization strategy. Compared with existing technologies, this invention has made significant progress in terms of technical comprehensiveness, system completeness, and practical reliability. At the same time, this invention has been verified by simulation experiments and real quantum hardware. The system can provide effective quantum acceleration during model training while still maintaining good noise robustness and resource efficiency. In particular, with the support of the adaptive quantum circuit optimization strategy and multi-level error mitigation mechanism, the system's performance on real quantum hardware is significantly improved compared with existing technologies. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0054] Figure 1 is a schematic diagram of the system of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] The core idea of ​​this invention is to construct a hierarchical and collaborative classical-quantum hybrid architecture, and to achieve deep collaboration between quantum computing and classical computing through dynamic allocation of quantum computing resources and adaptive quantum circuit optimization strategies, thereby enabling efficient model training tasks to be completed in various computing environments.

[0057] The specific functional modules of the model training and optimization system based on a classical-quantum hybrid architecture provided by this invention are described below with specific examples (as shown in Figure 1). The system functional modules include: a system configuration module, a data preprocessing and quantum state encoding module, and a quantum circuit optimization and error mitigation module;

[0058] The system configuration module is used to configure and schedule the resources required by the data preprocessing and quantum state encoding module and the quantum circuit optimization and error mitigation module;

[0059] The data preprocessing and quantum state encoding module uses classical computing devices to preprocess the raw data, and uses quantum computing devices to encode the preprocessed data into quantum states, embedding the quantum states into a high-dimensional quantum feature space.

[0060] The quantum circuit optimization and error mitigation module performs computational optimization and result optimization on the preprocessed data and quantum states, and finally combines the classical calculation results and quantum calculation results to output the optimal result.

[0061] I. System Configuration Module

[0062] The system configuration module includes a task scheduling layer, a computation execution layer, and a resource management layer;

[0063] The task scheduling layer is deployed on high-performance classical computing devices, using multi-core CPUs, GPU clusters, or distributed computing nodes. The task parsing unit uses natural language processing technology to parse user-submitted training tasks, identifying task type, data characteristics, and performance requirements. The resource allocation unit dynamically allocates classical and quantum computing resources based on task requirements and system resource status. The process scheduling unit uses directed acyclic graph (DAG) technology to manage the task execution process, ensuring the correct execution of dependencies between tasks.

[0064] The computation execution layer employs a heterogeneous computing architecture and intelligent decomposition algorithm. Based on the characteristics of the computation task, the load of classical computing devices, and the performance of quantum computing devices, it dynamically determines whether a subtask is suitable for classical processing or quantum acceleration. The classical computing devices include CPUs and GPUs, which are responsible for executing traditional computing tasks. The quantum computing devices are accessed through a quantum cloud platform or a local quantum processor, and support various quantum hardware computing platforms, including superconducting qubits and ion traps, through standard computing power access interfaces.

[0065] The resource management layer comprises a classical resource management unit and a quantum resource management unit. The classical resource management unit is responsible for managing classical computing resources, storage resources, and network resources, and uses distributed resource management technology to achieve unified management of large-scale classical computing resources. The quantum resource management unit is responsible for managing qubit resources, quantum gate resources, and quantum circuit compilation optimization, and uses quantum circuit compilation optimization algorithms to transform high-level quantum circuit descriptions into low-level instructions executable by quantum computing devices.

[0066] II. Data Preprocessing and Quantum State Encoding Module

[0067] The data preprocessing and quantum state encoding module includes a data preprocessing unit, a quantum state encoding unit, and a quantum feature mapping unit;

[0068] The data preprocessing unit uses classical computing equipment to preprocess the raw data;

[0069] The quantum state encoding unit uses a quantum state encoder to encode the preprocessed data into quantum states.

[0070] The quantum feature mapping unit embeds the quantum state into a high-dimensional quantum feature space after the data is encoded into a quantum state using quantum feature mapping technology.

[0071] 1) The data preprocessing unit uses classical computing equipment to preprocess the raw data, including:

[0072] Data cleaning: handling missing values, outliers, and duplicate values ​​to ensure data quality;

[0073] Feature engineering: Enhances feature representation capabilities by selecting relevant features, constructing new features, and transforming feature representations;

[0074] Data normalization: scaling data to a range suitable for quantum state encoding, such as [0,π] or [-π,π].

[0075] Dimensional Adjustment: The feature dimension is adjusted according to the number of qubits, and principal component analysis or an autoencoder is used to compress the feature dimension to 2. n Dimension, n is the number of qubits.

[0076] 2) The quantum state encoding unit uses a quantum state encoder to perform quantum state encoding on the preprocessed data, including:

[0077] An adaptive encoding method selection mechanism is introduced to automatically select the most suitable encoding method from the following options based on task requirements and data characteristics, thereby improving encoding efficiency:

[0078] Basis encoding: directly mapping the binary representation of classical data to the quantum ground state;

[0079] Amplitude encoding: Representing classical data as the amplitude of quantum states to achieve exponentially compressed storage;

[0080] Angle encoding: Maps classical data to the rotation angle of a quantum gate, suitable for continuous data;

[0081] Hamiltonian encoding: Encoding classical data into Hamiltonian parameters for use in quantum simulation tasks.

[0082] 3) After the data is encoded into a quantum state, the quantum feature mapping unit uses quantum feature mapping technology to embed the quantum state into a high-dimensional quantum feature space, including:

[0083] Quantum states are embedded into high-dimensional quantum feature spaces using custom quantum circuits or parameterized quantum circuits:

[0084] Custom quantum circuits: Achieving complex feature mappings by designing quantum circuit structures;

[0085] Parameterized quantum circuits: Using trainable parameterized quantum gate sequences, feature mappings are optimized by adjusting parameters.

[0086] III. Quantum Circuit Optimization and Error Mitigation Module

[0087] The quantum circuit optimization and error mitigation module includes an adaptive quantum circuit optimization unit and a quantum error mitigation unit;

[0088] The adaptive quantum circuit optimization unit adopts an adaptive quantum circuit optimization strategy to optimize quantum circuits based on task requirements and hardware characteristics.

[0089] The quantum error mitigation unit employs a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing.

[0090] 1) The adaptive quantum circuit optimization unit employs an adaptive quantum circuit optimization strategy to optimize quantum circuits based on task requirements and hardware characteristics, including:

[0091] Quantum circuit structure optimization: Select suitable circuit templates from the parameterized quantum circuit template library according to task requirements to balance circuit expressiveness and complexity;

[0092] Dynamic circuit pruning: By analyzing the importance of each quantum gate in a quantum circuit, redundant quantum gates that have little impact on quantum computing results are removed, thereby reducing the circuit depth;

[0093] Noise-adaptive compilation: Based on the noise characteristics and error rate of the quantum computing device, select a quantum gate decomposition scheme and a qubit mapping strategy with lower noise.

[0094] 2) The quantum error mitigation unit employs a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing, including:

[0095] Zero-noise extrapolation: The original quantum circuit is randomly compiled into a series of equivalent quantum circuits using random compilation technology. The noise level is adjusted using dynamic decoupling technology, and these quantum circuits are run under different noise levels. The quantum computing results at each noise level are collected, and the quantum computing results at zero noise are estimated using an extrapolation algorithm.

[0096] Measurement error mitigation: Construct a measurement error calibration matrix, measure the output distribution under all calculated ground-state inputs, and correct the actual measurement results using deconvolution or maximum likelihood estimation;

[0097] Probabilistic error elimination: The original quantum circuit is randomly compiled into a series of equivalent quantum circuits using random compilation technology. With the help of the quasi-probabilistic decomposition framework, the noiseless quantum operation is approximately expanded into a convex combination of noisy quantum operations under a specific noise model. The impact of noise on quantum information processing is analyzed, and the impact of noise is offset by weighted averaging of multiple experimental results.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model training and optimization system based on a classical-quantum hybrid architecture, characterized in that: It includes a system configuration module, a data preprocessing and quantum state encoding module, and a quantum circuit optimization and error mitigation module. The system configuration module is used to configure and schedule the resources required by the data preprocessing and quantum state encoding module and the quantum circuit optimization and error mitigation module. The data preprocessing and quantum state encoding module uses classical computing devices to preprocess the raw data, uses quantum computing devices to encode the preprocessed data into quantum states, and embeds the quantum states into a high-dimensional quantum feature space. quantum The circuit optimization and error mitigation module performs computational optimization and result optimization on the preprocessed data and quantum states, and finally combines the classical and quantum computation results to output the optimal result.

2. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 1, characterized in that: The system configuration module includes a task scheduling layer, a computation execution layer, and a resource management layer. The task scheduling layer is deployed on high-performance classical computing devices, employing multi-core CPUs, GPU clusters, or distributed computing nodes. The task parsing unit uses natural language processing technology to parse user-submitted training tasks, identifying task type, data characteristics, and performance requirements. The resource allocation unit dynamically allocates classical and quantum computing resources based on task requirements and system resource status. The process scheduling unit uses directed acyclic graph (DAG) technology to manage the task execution process, ensuring the correct execution of dependencies between tasks. The computation execution layer employs a heterogeneous computing architecture and intelligent decomposition algorithms to dynamically determine which subtasks are suitable for classical computing based on the characteristics of the computation task, the load of the classical computing device, and the performance of the quantum computing device. The system comprises a processing or quantum acceleration layer. Classical computing devices, including CPUs and GPUs, are responsible for performing traditional computing tasks. Quantum computing devices are accessed through quantum cloud platforms or local quantum processors, supporting various quantum hardware computing platforms, including superconducting qubits and ion traps, via standard computing power access interfaces. The resource management layer includes classical resource management units and quantum resource management units. The classical resource management unit manages classical computing resources, storage resources, and network resources, employing distributed resource management technology to achieve unified management of large-scale classical computing resources. The quantum resource management unit manages qubit resources, quantum gate resources, and quantum circuit compilation optimization, using quantum circuit compilation optimization algorithms to transform high-level quantum circuit descriptions into low-level instructions executable by quantum computing devices.

3. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 1, characterized in that: The data preprocessing and quantum state encoding module includes a data preprocessing unit, a quantum state encoding unit, and a quantum feature mapping unit; The data preprocessing unit uses classical computing equipment to preprocess the raw data; the quantum state encoding unit uses a quantum state encoder to encode the preprocessed data into quantum states. The quantum feature mapping unit embeds the quantum state into a high-dimensional quantum feature space after the data is encoded into a quantum state using quantum feature mapping technology.

4. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 3, characterized in that: The data preprocessing unit uses classical computing equipment to preprocess the raw data, including: data cleaning: handling missing values, outliers, and duplicate values ​​to ensure data quality; feature engineering: improving feature representation capabilities by selecting relevant features, constructing new features, and transforming feature representations; data normalization: scaling the data to a range suitable for quantum state encoding; and dimensionality adjustment: adjusting the feature dimension according to the number of qubits, and compressing the feature dimension to 2 using principal component analysis or an autoencoder. n Dimension, n is the number of qubits.

5. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 3, characterized in that: The quantum state encoding unit employs a quantum state encoder to encode the preprocessed data into quantum states. This includes: introducing an adaptive encoding method selection mechanism to automatically select the most suitable encoding method from the following options based on task requirements and data characteristics, thereby improving encoding efficiency: base encoding: directly mapping the binary representation of classical data to the quantum ground state; amplitude encoding: representing classical data as the amplitude of the quantum state, achieving exponential compression storage; angle encoding: mapping classical data to the rotation angle of the quantum gate, suitable for continuous data; Hamiltonian encoding: encoding classical data into the parameters of the Hamiltonian, used for quantum simulation tasks.

6. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 3, characterized in that: After the data is encoded into a quantum state, the quantum feature mapping unit uses quantum feature mapping technology to embed the quantum state into a high-dimensional quantum feature space, including: embedding the quantum state into the high-dimensional quantum feature space using a custom quantum circuit or a parameterized quantum circuit: custom quantum circuit: complex feature mapping is achieved by designing a quantum circuit structure; parameterized quantum circuit: feature mapping is optimized by adjusting parameters using a trainable parameterized quantum gate sequence.

7. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 1, characterized in that: The quantum circuit optimization and error mitigation module includes an adaptive quantum circuit optimization unit and a quantum error mitigation unit; The adaptive quantum circuit optimization unit employs an adaptive quantum circuit optimization strategy to optimize quantum circuits based on task requirements and hardware characteristics; the quantum error mitigation unit uses a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing.

8. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 7, characterized in that: The adaptive quantum circuit optimization unit employs an adaptive quantum circuit optimization strategy to optimize quantum circuits based on task requirements and hardware characteristics. This includes: quantum circuit structure optimization: selecting suitable circuit templates from a parametric quantum circuit template library according to task requirements to balance circuit expressiveness and complexity; dynamic circuit pruning: removing redundant quantum gates with minimal impact on quantum computing results by analyzing the importance of each quantum gate in the quantum circuit, thereby reducing circuit depth; and noise-adaptive compilation: selecting a less noisy quantum gate decomposition scheme and qubit mapping strategy based on the noise characteristics and error rate of the quantum computing device.

9. The model training and optimization system based on a classical-quantum hybrid architecture according to claim 7, characterized in that: The quantum error mitigation unit employs a multi-level error mitigation mechanism to reduce the impact of noise in quantum computing, including: zero-noise extrapolation: using random compilation technology to randomly compile the original quantum circuit into a series of equivalent quantum circuits, using dynamic decoupling technology to adjust the noise level, and running these quantum circuits under different noise levels, collecting the quantum computing results at each noise level, and using an extrapolation algorithm to estimate the quantum computing results under zero noise; measurement error mitigation: constructing a measurement error calibration matrix, measuring the output distribution under all computational ground state inputs, and using deconvolution or maximum likelihood estimation to correct the actual measurement results; probabilistic error elimination: using random compilation technology to randomly compile the original quantum circuit into a series of equivalent quantum circuits, using a quasi-probabilistic decomposition framework, approximating the noiseless quantum operations into a convex combination of noisy quantum operations under a specific noise model, analyzing the impact of noise on quantum information processing, and offsetting the impact of noise by weighted averaging of multiple experimental results.

Images